]> Huawei

China bill.wu@huawei.com

China Mobile

China zhouchengyjy@chinamobile.com

Telefonica

luismiguel.contrerasmurillo@telefonica.com

China Unicom

China hans29@chinaunicom.cn

Daejeon University

Korea yonggeun.hong@gmail.com

Network Management network digital twin large language model architecture network management A Network Digital Twin (NDT) provides a network emulation tool usable for different purposes such as scenario planning, impact analysis, and change management. Agentic AI enables dynamic goal-driven execution and adaptive behavior and closed-loop autonomy. By integrating a NDT into network management together with the Agentic AI, it allows the network management activities to take user intent or service requirements as input, automatically assess, model, and refine optimization strategies under realistic conditions but in a risk-free environment. Such environment that operates to meet these types of requirements is said to have AI-driven network operations. AI-driven network operations brings together existing technologies such as Agentic AI and NDT which may be seen as the use of a toolbox of existing components enhanced with a few new elements. This document describes an architecture for AI-driven network operations and shows how these components work together with NDT and Agentic AI capabilities. Discussion Venues Discussion of this document takes place on the Network Management mailing list (nmrg@irtf.org), which is archived at . Source for this draft and an issue tracker can be found at .

Introduction With the rapid expansion of network scale and the increasing demands on these networks, significant challenges arise with keeping device configurations consistent with the needs of services and rapidly adapting the network at scale. The intent based management has been first introduced in to address these challenges and necessitate of continuous network reconfiguration to better adapt to ever-changing service requirements. Since network changes are directly related to service operations, any successful change needs to not only ensure that new services are provisioned smoothly, but also that existing services are not affected and that no problems are introduced with the new configurations. On one hand, these changes might still require manual operation or iterative input from network Operators to provide operational guidance and information about the goals, purposes, and service instances that the network is to serve and refine the intent. On the other hand, pre-programmed automation scripts or policy templates may not be sufficient to realize the intent. Network operators are, therefore, increasingly cautious about making network changes. For example, they need to review the solution design as well as evaluate all change impacts, before making any change. Then, after the change, they need to perform dialing tests, monitor traffic, and manually check table entries. The Network Digital Twin (NDT) provides Network operators with a safe sandbox for testing complex configuration changes, software updates, or new protocols before they are deployed in production so that the risk of accidental outages or service disruptions can be greatly reduced. In addition, Agentic AI introduces disruptive paradigm to the network management which:

• Allows declarative intent interpretation and refinement with multi-step task management,
• Enables dynamic goal-driven execution and adaptive behavior instead of stepwise static instruction,
• And enables closed-loop autonomy across layer or across domain instead of deterministic closed-loop automation.

Integrating a NDT into network management together with Agentic AI, it allows network management activities to dynamically adapt to customer needs, network changes, as well as to automatically assess, model, and refine optimization strategies under realistic conditions but in a risk-free environment. An environment that operates to meet these types of requirements is said to have AI-driven network operations. AI-driven network operations provide the following capabilities to applications by coordinating the components that operate and manage the network:

• Service intent and service assurance work together to ensure that the network change or network optimization aligns with business goals and that the services provided meet the agreed-upon Service Level Agreements (SLAs).
• Provide network capacity planning and ensure that the network has sufficient capacity , resources, and infrastructure to meet current and future demands.
• Provide simulation on fault scenarios, formulate recovery plans, and verify whether the plans are applicable and effective so that the service will not be affected during disaster recovery drill.
• Support fault and risk detection and provide network health check and network risk check.
• Model the network configuration change and use a virtual topology model to test network changes and assess the effect of the network configuration changes on the network.
• Model the protocol operations and interactions among devices in the network and simulate specific networking protocols such as IS-IS, OSPF, BGP, SR, etc to understand how they perform under different conditions.
• Model traffic flow across the network, including traffic generation, flow control, routing, and congestion control and evaluate traffic's impact on network performance.
• Support generation of repair solutions for potential network risks and provide verification on the repair solution in seconds, including loop, address conflict, and security policy conflict.
• Model the performance characteristics of each flow according to traffic conditions change by using a virtual topology in order to understand which flows and network portions can be potentially affected by performance degradation and act to prevent it

This document describes an architecture for AI-driven network operations, showing how these components work together with NDT and AI capabilities.

Conventions and Definitions The document uses the following definitions and acronyms defined in :

• Network Digital Twin (NDT)
• Artificial Intelligence (AI)

The following acronyms are used throughout this document:

• Generative Artificial Intelligence (Gen-AI)
• Large Language Model (LLM)
• Retrieval-Augmented Generation (RAG)
• Agentic AI
• Remote Code Execution (RCE)

Besides, this document defines the following terminology:

Network AI Agent:

Network AI Agent is a persistent, long-lived autonomous entity with continuous awareness of its network environment, capable of conducting analysis, making decisions, and executing actions with specific intent based on its knowledge representation to achieve a set of service goals . In addition, it is able of planning the tasks and decompose the tasks into several sub-tasks and coordinate with Task AI Agent for these sub-tasks.

Task AI Agent:

Task AI Agent is a domain-specialized execution entity responsible for carrying out specific subtasks delegated by the Network AI Agent. Depending on operational requirements, a Task AI Agent operates in one of two following modes:

Ephemeral Task AI Agent:

Created on demand by a Network AI Agent for a specific, scoped subtask (e.g., executing a single troubleshooting action). It exists only for the duration of that task and is destroyed upon completion, avoiding registry bloat and lifecycle management overhead.

Persistent Task AI Agent:

A long-lived, pre-provisioned entity dedicated to continuous, specialized operational needs (e.g., long-term anomaly detection) that require continuous network state tracking.

Autonomy:

The ability to operate under uncertainty and adapt to changing objectives without explicit external intervention.

Introduction of Concepts

Generative AI and Agentic AI The integration of AI into network operations has marked a significant leap forward in the pursuit of network automation and intelligence, while generative AI further enhances the role of AI-driven network operations and management. Generative AI is a subfield of AI that uses generative models such as Large Language Models (LLMs) to generate new and original content such as text, images, videos, or other forms of data with the capability to adapt and make decisions to achieve specific goals. Agentic AI refers to the broader category of AI systems that exhibit "agency"—the ability to act independently and iteratively to perform tasks without constant human prompting . In the context of network operations and management, Network AI agents are increasingly being designed to interact with physical world and act upon it based on tools and perform network management tasks such as understanding user intent, generating network configurations, diagnosing and resolving network incidents . Meanwhile, other SDOs also try to define terms related to Network AI agent in the context of network operations and management, e.g., TM Forum defines Autonomous Agent in as one of AN (Autonomous Network) Terminologies.

Network Digital Twin The NDT is a digital representation that is used in the context of network. The concept and architecture of the NDT are specified in . Three core functional components which includes Data Repository component, a Service Mapping Models component, and an NDT Management component are introduced to characterize the NDT and its reference architecture. The NDT is widely recognized to be useful as an advanced platform for network emulation, serving as a tool for scenario planning, impact analysis, and change management. By delivering applications requests to the NDT through standardized interfaces (see ), the NDT exposes the various capabilities to network applications.

Characteristics of AI-driven Network Operations AIOPS was first defined by Gartner in 2016, combining "artificial intelligence" and "IT operations" to describe the application of AI and machine learning to enhance IT operations. However there is no unified definition for characteristic of "AI-driven network operations" within the networking industry. Referring to the characteristics of AIOPS in IT field and the characteristics of networking itself, this document introduces six key elements (i.e., awareness, decision, analysis, execution, intent and knowledge) to characterize the AI-driven network operation and its use, as shown in . They together form a close-loop of network operation and management.

Six Key Elements to Characterize AI-driven network operation

• Intent:

  Intent is defined as a set of operational goals and outcomes defined in a declarative manner without specifying how to achieve or implement them in . The Network AI Agent must accurately interpret and understand the user's high-level business or operational objectives, this involves translating declarative requirements into specific network instructions, e.g., configurations.

• Knowledge:

  The Network AI agent relies on a knowledge base that includes network policies, historical data, expert experience, extra-system experience (updates to LLMs/their implied ‘knowledge bases’) and Manually or semi-manually entered knowledge, e.g., new equipment spec sheets,best practices in product manual. The knowledge is used to inform its analysis, decision-making, and execution processes. Over time, the Network AI agent can expand its knowledge through machine learning, incorporating new data and experiences to improve its performance. For example, it learns which configurations are optimal for specific scenarios or how to respond most effectively to particular types of network incidents .

• Analysis:

  The Network AI agent continuously analyzes vast amounts of network data from various sources, including network telemetry and external feeds, and identify the gap between user intent and the existing network status. By integrating NDT with Network AI agent and leveraging machine learning and other data analytics techniques, it also identifies network fault, problem, incident, anomaly and perform data-driven intelligent analysis such as service impact analysis, and so on. Their distinction is further discussed in .

• Decision:

  Based on the intent and network analysis, AI makes informed decisions. By integrating NDT and AI, the intelligence decisions making can be realized. These decisions could involve dynamically adjusting network parameters, e.g., rerouting traffic to avoid congestion. The decision-making process is driven by predefined policies, real-time data analysis, and AI models (e.g., LLMs) that enable the Network AI agent to choose the best course of action to meet the specified intent. Network AI agent may also verify the correctness of the decision outcome by performing some network simulation or validation process.

• Awareness:

  Awareness is achieved through real-time monitoring and data collection. The Network AI agent maintains a comprehensive visibility of the network, enabling it to make context-aware decisions. Network operators can also use the awareness understand the exact cause of specific network issues and achieve closed-loop decision-making.

• Execution:

  Once a decision is made, the Network AI agent executes the necessary actions to implement it. This could involve, e.g., sending configuration to network controllers or network devices through NETCONF/RESTCONF protocols. The execution is carried out in a controlled and precise manner to ensure that the network behaves as intended without causing disruptions. The Network AI agent also verifies that the executed actions have the desired effect and makes the proper adjustments if needed.

Architecture Design

Overall Architecture provides the overall architecture for integrating Network Digital Twin and Network AI Agent System. The components and functional interfaces are discussed in and , respectively. The use cases described in show how different components are used selectively to provide different services. It is important to understand that the relationships and interfaces shown between components in this figure are illustrative of some of the common or likely interactions; however, this figure does not preclude other interfaces and relationships as necessary to realize specific functionality.

An Architecture for Integrating Network AI Agent with Network Digital Twin +--------------+| | || | | | | | ||Security&Trust|| | ||+------v---+ +-----v----+ +--------v------+ v | |+--------------+| | |||Task Agent<->|Task Agent<->|Function Module|...| |+--------------+| | ||+----------+ +----------+ +---------------+ | ||Observability,|| | || | || C & I || | |+-----------^-----------------------^------------+ |+--------------+| | | | | |+--------------+| | |+-----------v----------+ | ||Knowledge Base|| | || Network Digital Twin | | |+--------------+| | |+-----------^----------+ | +----------------+ | | | | | |+-----------------------------------v---------------------------------+ | ||Physical Network | | || +---------+ +------------------+ +---------+ | | || | | | | | | | | || | NE | |NE(lightweight AI)| ... | NE | | | || | | | | | | | | || +---------+ +------------------+ +---------+ | | |+----------------+----------------------------------------------------+ | +------------------------------------------------------------------------+ ]]>

Functional Components This section describes the functional components shown as boxes in . The interactions between those components, the functional interfaces, are described in .

Network Applications Various network applications at the service level can effectively run over a AI-driven network operation platform to implement either conventional or innovative network operations, with low cost and less service impact on real networks. A network application may be a software tool that a user uses to make requests to the network to set up specific services such as end-to-end connections or scheduled bandwidth reservations or NOC Application /Service AI Agent Application that is responsible for monitoring, managing, and maintaining the health, performance, and availability of complex networks. Network applications make requests that need to be addressed by the AI-driven network operation. Such requests are exchanged through a either traditional northbound intent interface (e.g., Restful API), or A2A style intent interface with Bi-directional interaction support and Multi-step dialogue for intent refinement (e.g., Natural Language Programming Interface(NLPI),A2A, A2A-T), so that they can be applied by Hybrid Agent System to either interact with traditional deterministic systems (OSS/BSS, orchestration) or Autonomous service AI agents acting as consumers of the Hybrid Agent System.

Autonomous Domain An autonomous domain is a self-governing network that integrates NDT and AI-driven capabilities to achieve autonomous network management. It comprises the following sub-components.

Hybrid Agent System Hybrid agent system acts as the smart brain of the Autonomous Domain, which is responsible for conducting AI-based analysis and making decisions regarding network management operations. It usually comprises a Network AI Agent and multiple ephemeral or persistent task agents or function modules (e.g., Agent skills, network management functional APIs or tools). It is worth mentioning that function modules could be the functions and services of existing network management systems, it offers a set of stateless APIs and tools that can be consumed by agents. In some simplified scenarios, the system may consist of only one single AI Agent equipped with multiple function modules, e.g., skills and toolsets that integrates both network-level coordination and task-specific execution. The Network AI Agent may coordinate cross-task-agent collaboration, aligns tasks with user intent, and supervises the execution of each task agent or function module. And task agents and function modules are designed to perform specific functionalities, they could be scenario-oriented and classified according to the function they perform. The hybrid agent system can adapt to new circumstances through access to evolving knowledge and reasoning, planning. It leverages the inference of LLM, the simulation of NDT, and the contextual and domain-specific knowledge provided by Knowledge Base to accomplish specific network operation task.

Agent Fabric The Agent Fabric acts as the unified management hub and internal communication infrastructure for the hybrid agent system, including agent registration/discovery, authentication, observability, and knowledge base. With the introduction of the agent fabric, consistent observability, traceability, control and security can be provided. Specifically, within a single autonomous domain, the Agent Fabric provides localized agent management and communication facilitation. When extending to multi-domain scenarios, the Agent Fabric is leveraged across multiple domains to enable cross-domain agent registration/discovery, trust establishment, and execution traceability.

Registration AI Agents need to first discover each other and understand their capabilities to collaborate. Agent Registration manages the process by which new agents could join the system, making them discoverable and available. It supports the unified registration of all AI agents across the autonomous domains. Note that ephemeral task agents is hidden behind the domain AI agent and short lived created on demand so that they will not be registered in the same way as network AI gent in each autonomous domain, instead the task agents capabilities can be registered in the agent registry. Each Agent instance submits its own metadata information including URI, supported authentication methods, and capabilities to the Agent Registry. And the consumer Agent (e.g., the Network AI Agent or task agent) could query or subscribe to the Agent Registry to find appropriate Agents for task execution. Agent skills , introduced by Anthropic, provides a new way for Agents to improve how they perform specific tasks through folds that include instructions, scripts, and resources that are only loaded when needed. Skills are folds containing a "skill.md" file, Registration component may also work as a skill hub that enables Agents to query directories for required skills. implements Agent Registration by providing the Agent Card mechanism to ensure Agents from different vendors can register and discover other Agents they need.

Security & Trust The security component enforces trusted inter-Agent communication by verifying the identity of AI Agents and enforcing security policies throughout their interaction. It provides unified security functionalities for all AI Agents within the autonomous domain, including those residing on network devices. Some existing authentication methods such as OAuth 2.0, allow to issue each AI Agent its own authentication credentials to establish trusted communication. Standardized protocols like TLS (Transport Level Security) could be leveraged to protect sensitive data exchanged between AI Agents. It is also worth noting that once authenticated, authorization defines the specific tools and data an agent can access, which often using a Least Privilege access control method. It is also recommended to log every Agent decision and tooling call to maintain audit trail.

Observability, C & I Observability, Control and Intervention component provides monitoring, policy enforcement, and real-time intervention capabilities for all AI agents within the autonomous domain, enabling network operators to gain deep insights into agent behaviors and maintain determinism over autonomous network operations. OpenTelemetry(OTel) has emerged as a vendor-neutral approach for AI Agent observability. It provides a consistent instrumentation layer, semantic conventions, and telemetry export mechanisms, ensuring end‑to‑end visibility across network AI agents, task AI agents, LLM inference, and tool/skill executions. By leveraging Semantic Conventions for GenAI , the system unifies the collection of LLM inference metadata, token consumption, reasoning processes, and skill invocation events, etc. This allows comprehensive tracking of agent runtime status, cross-component interaction flows, and abnormal behaviors in both multi-agent collaboration scenarios and simplified single-agent deployments with integrated multiple skills and tools. Beyond agent oversight, this component also incorporates control and runtime intervention mechanisms to guarantee agent operational compliance:

• Control:

  A prevention mechanism to constrain Agent behavior within operational boundaries . This includes predefining the agent's behavior scopes, operational constraints, and security baselines so that harmful or unauthorized actions are difficult or impossible to execute.

• Intervention:

  A reactive and emergency actions that enable the capability to pause, terminate, correct, or rollback agent execution (e.g., during LLM reasoning or tool invocation) at any point for any reason . Intervention provides humans with the ability to intervene or take control of agent behaviors that is not anticipated by control mechanism.

Knowledge Base The Knowledge Base serves as a crucial repository of information within the architecture. It enables the injection of expert knowledge and chain of thoughts, provides the necessary knowledge and memory that helps Agents make more accurate and practive context-aware decisions. It also helps mitigate the hallucination problems that can arise in large-scale models, which enhances the accuracy of task execution. Additionally, the Knowledge Base plays a key role in providing the data needed for techniques like Retrieval-Augmented Generation (RAG), which further boosts the system's ability to generate reliable and relevant outputs.

Network Digital Twin A NDT provides an enhanced and optimized solution in the face of increasing network and business types, scale, and complexity. It simulates the behavior, performance, and characteristics of the actual network, which could help in validation and testing scenarios, analyzing and predicting network behavior without affecting the real physical network. As described in , the core functional components of a NDT includes Data Repository, Service Mapping Models, and a NDT Management component. The NDT collects the real-time operational and instrumentation data from network through the appropriate real network-facing input interfaces, and it delivers NDT services through appropriate application-facing output interfaces, which is the interfaces to Network AI Agent(s) in .

Physical Network This is the actual hardware and infrastructure that makes up the network, which includes a set of network devices and wiring. In a physical network, Network Elements (NEs) with Lightweight AI or AI Agent may also achieve some local close loop without relying on human intervention. It is also possible for Lightweight AI or AI Agent to coordinate with other AI Agent(s) to enhance the automation and efficiency of network operations. The Network Lightweight AI models could be trained, validated, deployed, and executed on Network Elements, and further refined (e.g., model re-training) through monitoring and continuous optimization based on feedback from LLM.

Functional Interfaces This section describes the interfaces between functional components that might be externalized in an implementation allowing the components to be distributed across platforms. Where existing protocols might provide all or most of the necessary capabilities, they are noted. As noted in , it is important to understand that the relationships and interfaces shown between components in are illustrative of some of the common or likely interactions; however, this figure and the descriptions in the subsections below do not preclude other interfaces and relationships as necessary to realize specific functionality. Thus, some of the interfaces described below might not be visible as specific relationships in , but they can nevertheless exist.

Human Oversight (Human on the Loop) The architecture allows human experts to monitor, guide, approve, or intervene in the AI-driven network operations. Human may provide guidance and make critical decisions when necessary. By involving human in the process, the architecture can leverage their insights and experience, ensuring AI actions align with organizational goals. Human on the loop is also helpful to provide a safeguard for complex or sensitive decisions, where human judgement is essential to avoid potential errors or ethical dilemmas. This typically uses natural language as the primary mode of interaction, a chat platform that allows for conversational interaction with AI Agents can be leveraged. In some scenarios, operators may use structured format for strategy injection via workflows. Protocols like A2A , and RESTful API can be leveraged.

Application to Network AI Agent Interface (Intent based Management) Intent based Network Management helps in delivering application requests as service intent to the AI-driven network operation platform and allowing expose the various platform capabilities to network applications. Standardized protocols and interfaces facilitate smooth communication between applications and AI-driven network operation platform and ensures different systems from various vendors can work together seamlessly to support intent-driven network management. One example of the interface between Network applications (e.g., Service AI Agent) and Network AI Agent can adopt IG1453 Agent to Agent Protocol for Telecoms (A2A-T) specified by TM Forum.

Network AI Agent to Task AI Agent Interface (Single Autonomous Domain) This interface governs the coordination and task delegation within the Hybrid Agent System of a single Autonomous Domain. The Network AI Agent, acting as the principal coordinator, uses this interface to decompose high-level goals into specific tasks and assign them to specialized Task Agents (e.g., for configuration generation or fault diagnosis). It facilitates communication for task assignment, progress monitoring, and result aggregation. This coordination can be implemented using protocols like .

Network AI Agent to Network AI Agent Interface (Cross Autonomous Domain) This interface enables collaboration and information exchange between Network AI Agents residing in different Autonomous Domains. It is essential for scenarios requiring end-to-end service assurance or coordinated optimization across multi-domain networks. Through this interface, Network AI Agents can negotiate resource allocation, share summarized domain-specific insights (while preserving detail isolation for privacy and scalability), and coordinate actions to fulfill cross-domain objectives. Standardized protocols like A2A-T , designed for agent interoperability in telecommunication area, are candidate technologies for implementing this cross-domain interface, ensuring secure and reliable interaction between autonomous systems from different administrative domains.

Hybrid Agent System to Agent Fabric Interface (Agent Lifecycle Management) The interface between Hybrid Agent System and Agent Fabric serves as the management bridge which encompasses a set of services designed to manage the lifecycle of AI agents, and collaborative capabilities of AI Agents such as Registration, Security and Trust, Observability.

• Registration Registration handles Agent onboarding, lifecycle tracking (e.g., heartbeat monitoring, status updates), and capability-based Agent discovery. Interfaces like RESTful APIs with structural schema for AI Agents metadata description could be leveraged. Protocols like A2A Agent card mechanism may also be used to ensure interoperability among different Agent vendors. It is also worth noting that message queue mechanisms such as Kafka could also be a candidate interface for asynchronous communications for agent registration and discovery. This interface also supports skill management, including skill query, loading, unloading, and version control, enabling agents to dynamically obtain and use skills provided by the Registration component.
• Security and Trust Security and Trust ensures trusted inter-Agent communication by verifying the identity of AI Agents and enforcing security policies throughout their interaction. Protocols like Transport Layer Security (TLS) could be leveraged for in-transit data Protection. While OAuth 2.0 and OpenID Connect are increasingly used to authenticate AI Agents.
• Observability The observability component enables the network operators to gain deep insights of agent behaviors. It collects audit logs, metrics and traces of each agent and provides visibility of task process. In addition, to make the behavior of Hybrid Agent System declarative and traceable, it is essential to have the capability to distinguish whether an operation action is taken by agents automatically or by human operators. For example, in the fault management scenarios that need on-site repairing, the demarcation and locating task could be accomplished by agents, and the repair solution implementation and verification task is accomplished by human operators, maybe with the help of intelligent copilot. The traceable logs are essential for both clarifying the duty boundary between agents and human operators, and also for evaluating the effectiveness of the agentic system.

The interface between AI Agent and Knowledge Base is specified in .

Hybrid Agent System to Network Digital Twin Interface The interface between Hybrid Agent System and NDT are the application-facing interface as defined in . Furthermore, the Model Context Protocol (MCP) can be leveraged to standardize this interaction, enabling the NDT to expose its simulation and analysis capabilities as a set of discoverable "tools" that the AI Agent can dynamically invoke. This MCP-based approach facilitates seamless integration and richer contextual exchange between the Agent and the NDT.

Hybrid Agent System to Knowledge Base Interface Knowledge Base service provides contextual data and insights to enhance the decision-making accuracy of the Hybrid Agent System. Interfaces such as Cypher or SPARQL with schema-defined data models (e.g., LPG or RDF for knowledge representation) allow efficient retrieval and updates. Other high-throughput interfaces such as gRPC or RESTful API can be the candidate for synchronous semantic search queries. For large-scale knowledge operations, asynchronous data message systems (e.g., Kafka) can also be employed for data ingestion and real-time knowledge synchronization across distributed Agents. Additionally, MCP could also serve as a standardized interface for AI Agents to dynamically access and utilize a wide range of tools and data sources provided by the Knowledge Base. It enables the Knowledge Base to expose contextual information, expert rules, and external data as "tools" that Agents can invoke, significantly enhancing their reasoning and problem-solving capabilities.

Hybrid Agent System to Physical Network Interface The interface between the hybrid agent system and the physical network is established via its internal function modules interacting with network elements (NEs). AI agent can invoke function modules to perform specific network operations.

Data Collection Data Collection interface is responsible for gathering data from the physical network through various different tools and methods (e.g., IPFIX , YANG-push ,, and BMP ). It collects various types of network data including configuration data, operational data, network topology, routing data, logs, and trace on management plane, control plane, and forwarding plane as needed. The collected data is fed into the NDT and hybrid agent system to provide with up-to-date information about the current state of the physical network.

Configuration Once network decisions are made and confirmed, the Hybrid Agent System performs specific actions to the physical network, e.g., modify specific configuration through protocols like NETCONF and RESTCONF , which utilize YANG as the data modeling language. It is the component that makes the planned control and management changes a reality in the real physical network.

Lightweight AI and Large AI Model Collaboration Interface Collaboration between small AI model and large AI model is also designed to be supported by this interface. In the past, we only support AI and machine learning technologies at the network level, e.g., we can use collected various different network data to provide network analysis and generate network insight. With more intelligence introduced into the network element, more GPU/NPU resource can be allocated for AI inference, this make collaboration between large AI model and small AI model possible. Large AI models can provide basic logical reasoning and generalized analytical decision-making capabilities While specialized small AI models can provide efficient problem-solving capabilities in specialized areas. The synergy between the two allows the AI agent to combine both multitasking generalization capabilities and domain expertise, thus minimizing the reliance on human intervention in the network management process. On one hand, we can use accumulated field engineering expertise to train large AI model into one foundation model for fault management AI agent, On the other hand, we can deploy small AI model, leverage hardware resource or chipset resource in the intelligent network element to collect more fine granularity data or provide local processing for Collected data and summary report generation, Trend prediction, etc. When small AI model is outdated and unable to detect specific applications or security risk, these specific applications and security risk information can be collected by network analytics platform to retrain this small AI model and re-deploy it in the same network element when this small AI model has been trained to work correctly to detect applications or security risk. With collaboration between large AI model and small AI model, we can allow Network AI Agent within the Network controller interact with network element and has more quick response to network change. This collaboration, facilitated by APIs or agent communication protocols like A2A , combines the generalization power of large models with the efficiency and low-latency of specialized small models, leading to quicker and more context-aware responses to network change.

Feedback-driven Improvement Interface The architecture should incorporate mechanism for continuous improvement based on feedback. This includes collecting data on AI decisions, network performance, and user feedback to identify areas for enhancement. By analyzing the feedback, the system can adapt and optimize its operations over time, leading to better performance and more accurate decision-making. For example, if a Network AI Agent fails to accurately identify the exact cause of a network incident, the relevant records can be submitted as negative samples to the LLM which provides inference services, this allows the LLM to be trained on these negative samples for optimization. This interface is implemented through a combination of system interfaces that collect, process, and apply feedback. Operational feedback—including the outcomes of AI decisions, network state metrics—is collected as structured data via system logging streams (e.g., in JSON format) and message queues (e.g., Kafka). This data is then consumed by analytics components and machine learning platforms through APIs (e.g., RESTful, gRPC) to refine AI models, for instance, by using failure records as negative samples for fine-tuning. Subsequently, optimized models and updated knowledge are deployed back into the runtime system via model serving and configuration management interfaces, closing the improvement loop.

AI-driven Network Operations: Relationship Between Characteristics and Functional Components The architecture in provides a concrete implementation framework to realize the six key characteristics of AI-driven network operations described in . Each characteristic is directly supported by specific functional components within the Autonomous Domain. The following clarifies how the architecture operationalizes these characteristics:

• Intent:

  The Network Application​s Layer conveys a high-level user intent via northbound interfaces. The Network AI Agent​ interprets this intent and translates it into actionable network operation tasks to each task Agent.

• Knowledge:

  The Knowledge Base​ in Agent Fabric serves as the central repository for domain-specific knowledge, expert rules, and historical data. It provides the necessary context and long/short memory to support accurate decision-making bytask Agents.

• Analysis:

  The AI Agent​ in Hybrid Agent System performs intelligent analysis using data and tools. It leverages the NDT to simulate and validate scenarios, enabling data-driven insights and gap analysis between intent and current network state.

• Decision:

  The AI Agent in Hybrid Agent System​ makes informed decisions based on its analysis results. It utilizes the NDT for risk-free validation before finalizing decisions. The decision may be sent to human operators for confirmation before actions are taken.

• Awareness:

  The AI Agent in Hybrid Agent System gathers data from the Physical Network, it may also fetch data from the NDT​ which maintains a dynamic, virtual representation of Physical Network. Together, they provide comprehensive network visibility and context-aware awareness.

• Execution:

  The AI Agent in Hybrid Agent System implements validated decisions by applying configurations or control actions to the Physical Network via southbound interfaces such as NETCONF, RESTCONF, or MCP.

AI-driven Network Operations: A collection of Use Cases Network AI Agent could help in the following phases which are usually mentioned in network management:

• Network Planning and Design: includes the understanding of user intent, generation of solutions, and simulation for decision-making.
• Service Deployment: includes the construction of the physical network, as well as intent understanding, pre-deployment simulation, automated configuration, post-deployment validation, and other capabilities to enhance the efficiency and accuracy of network configuration for service deployment.
• Network Monitoring and Troubleshooting: includes intent monitoring, issues identification, solution generation, evaluation and decision-making, solution implementation, and service validation.
• Network Change and Optimization: involves the design, evaluation, decision-making, implementation, and validation of network configuration changes or optimizations to improve network operation efficiency.

In all phases and use cases, after the Agent performs specific action, it always continuously monitors the network by data collection. Based on the result of network running analysis and user explicit feedback, it may adjust and optimize the management strategy if necessary.

Multi-Agent Collaboration on Network Configuration Change Network configuration changes are needed in scenarios such as optimizing network or service performance, provisioning new network services, or resolving network incidents/faults.

Intent Based Network Configuration Change Usage Example AI Agent | | | +-----+------+ |Configuration +---------+ |Change Intent | | +-------V---------+ | | |Network AI Agent | | | | Goal | | | +-------| Task-1,Task-2 +------------------> | | | ... Task-n +-------+ | Network | | +-------+---------+ | | Digital | | | | | Twin | | | | | | +-----V------+ +-----V-------+ +------V----+ | | | Config | | Config | | Config <-----> | | Generation | | Distribution| |Validation | | | | Task Agent | | Task Agent | | Task Agent| | | +------------+ +-------------+ +-----^-----+ +---------+ | +------V------+ | Resource | | Allocation | | Task Agent | +-------------+ ]]>

Network configuration change leveraging Network AI Agent and NDT may experience the following typical steps:

Step 1:

The network operator inputs the intent of network configuration change into the Network AI Agent using natural language. The network operator may explain the objectives and requirements of the changes.

Step 2:

Network AI Agent first verifies the identity of the user requesting the change and checks the user's permissions to make certain types of network changes against predefined rules or policies. It then understands and parses the initial intent of the request, by leveraging the powerful knowledge and reasoning capabilities of LLM and decompose the tasks into configuration generation task, configuration distribution task, configuration validation task and assign to corresponding task agents. Configuration generation Task Agent first generates initial suggestions for specific network configuration update, which may include multiple possible network configuration change plans if possible.

Step 3:

Network AI Agent further communicates with the Configuration Validation task agent and NDT to validate the suggested configuration change, including the syntax and semantics of the configuration, verification of affected application and resources. The NDT may generate a report indicating the validation result, and suggested configuration fix when the validation fails after network simulation leveraging the current physical network operational state.

Step 4:

Network AI Agent may generate a configuration change plan and submit to the network operator for approval. Based on the feedback from the operator, Network AI Agent then further decides whether to optimize the change plan or deliver the plan to the Configuration Distribution task agent to conduct the physical network configuration change. The configuration distribution task agent may further communicate with resource allocation task agent to obtain network resource (e.g.,vlan, IP subnet) allocated by resource allocation task agent.

Multi-Agent Collaboration on Network Troubleshooting

Intent based Network Troubleshooting Usage Example | | ....Task-n | | | +--------+--------+ +---------+ | |------------------+--+-----------+------------+ +---+----------+ +-----+-----+ +-----+----+ +-----+------+ | Fault | | Fault | | Fault | | Fault | |Identification| | Diagnosis | | Repair | | Prediction | | Task Agent | |Task Agent | |Task Agent| | Task Agent | +--------------+ +-----------+ +----------+ +------------+ ]]>

The network operator inputs the intent of network troubleshooting into the Network AI Agent using natural language. Network AI Agent could plan and decompose network troubleshooting tasks and coordinate with fault identification task agent, fault diagnosis task agent, fault repair task agent and fault prediction task agent to assist in network troubleshooting in the following significant aspects:

• Fault Identification:

  Network AI Agent coordinates with fault identification task agent continuously monitors and aggregates data from various sources, the comprehensive data collection provides a holistic view of the network operational state. By analyzing the real-time data, fault identification task Agent could detect network anomalies swiftly, which enables the prompt identification of potential issues before they escalate into major faults, minimizing downtime or service disruptions. In some cases, the Lightweight AI located in the Network Element may handle some simple fault identification tasks (e.g., optical module fault automatic identification) to enhance the awareness, while the fault identification task agent and LLM could leverage their powerful processing capabilities to analyze the time-domain data collected from the optical module.

• Fault Diagnosis:

  Once a fault is identified, Network AI Agent coordinate with fault diagnosis task agent to delve into diagnosing the exact cause, fault diagnosis task agent may also invoke some existing operations such as "incident-diagnose" RPC defined in . By correlating symptoms and/or applying AI models trained on historical data, fault diagnosis task agent can narrow down the potential causes and pinpoint the exact cause, which accelerates the diagnosis process and reduces the time needed to address the issue.

• Fault Repair:

  After diagnosing the fault, Network AI Agent can coordinate with fault repair task agent to generate targeted repair solutions. These solutions range from specific configuration adjustments to more complex fixes (e.g., hardware replacement). Fault Repair task Agent would also communicate with the NDT to simulate the proposed repair solutions and get feedback from the NDT. In advanced setups, fault repair task agent may automatically execute these repairs, ensuring quick restoration of normal operations and enhancing the overall reliability and efficiency of network management. But the fault repair task agent may also first present the fault details and repair advice to the network operator for review, and proceed to carry out the repair task once it is confirmed.

• Fault Prediction

  As an advanced enhancement of fault management capabilities, fault prediction aims to reduce network risks through proactive management that prevents problems before they occur. Before a fault actually occurs, the fault prediction task agent can coordinate with NDT to construct a dynamic simulation model by collecting real-time multi-dimensional operational state data, including network topology, traffic load, and device performance indicators. Based on the network data, the fault prediction task agent uses large models and machine learning algorithms (such as time-series prediction models and anomaly detection models) to reason and analyze potential faults—for example, predicting the risk of physical link interruption based on optical cable signal attenuation data. Furthermore, the fault prediction task Agent generates recommended operations to avoid faults and validates them through simulation in the NDT, thereby achieving predictive maintenance of the network.

Multi-Agent Collaboration on Network Optimization

Intent based Network Optimization Usage Example | | ....Task-n | | | +--------+--------+ +---------+ | +----------+----------------+ | | +-------+------+ +-------+-------+ | Optimization | | Optimization | | Generation | | Distribution | | Task Agent | | Task Agent | +--------------+ +---------------+ ]]>

Network optimization is often introduced due to the Network AI Agent's awareness of some potential network faults or anomalies through continuously monitoring of network operational state, e.g., AI models may predicts network congestion by analyzing historical and real-time network traffic data. It may also be triggered by the network operator actively inputting the network optimization intent. Based on the analysis of network data and user's intent (if any), Network AI Agent collaborate with Optimization Solution Generation Task Agent to propose network optimization strategies. For instance, once the network congestion sometime in the future is predicted, it may proactively optimize the network configuration, or suggest scaling up to meet specific demands. Before the network optimization is conducted, Network AI Agent coordinates with the NDT to implement and evaluate the optimization solution using the NDT platform. This may need repeated trials and validations based on specific evaluation criteria, before the optimal strategy could be selected. Network AI Agent may also first present the suggested network optimization solution to the network operator for review, and apply it to the physical network through optimization solution distribution task agent after obtaining approval from the network operator.

Network level Energy Efficiency Management in the IP + Optical network

Intent based Network level Energy Efficiency Management Usage Example

Network level Energy Efficiency refers to a set of processes used to discover an inventory of capabilities, use specific metrics to monitor and assess energy consumption of the entire IP+Optical network , operate, and control the use of available energy in an optimized manner while achieving the network’s functional and performance requirements by improving overall network utilization. Multi-Domain AI Agent can work together with network AI Agent in each autonomous domain to allow network operators not only see real time energy consumption in the network devices of large scale network through interaction with the GREEN Network AI Agent, but also allow them see o which network devices enable energy saving, which devices not, which are legacy ones, o The total energy consumption changing trend over the time of the day, for all network devices, o Energy efficiency changing trend over the time of the day for the whole network. On the other hand, With the end to end observability to energy consumption statistics data and energy efficiency statistics data, the Network AI Agent in each autonomous domain can collaborate with NDT to know which part of the network need to be adjusted or optimized based on network status change.

Network Security Drills (Human on the Loop)

Intent based Network Security Drill Usage Example |Observability |<-> Inject | | ....Task-n | || || \---------/ +-------+---------+ |+--------------+| | | | | +---------^------+ | | +-------V------ ---------------V----------------+ | Dynamic Attack and Defense Verification System| | Based on Network Digital Twin | | +---------+ +---------+ | | | Dynamic | | Dynamic | | | | Security| |Security | | | | Attack | |Defense | | | | Task | | Task | | | | Agent | | Agent | | | +---------+ +---------+ | +--------------------^--------------------------+ |Data Collection +--------------------+--------------------------+ | Network Infrastructure | +-----------------------------------------------+ ]]>

The human operator can work together with the Network AI Agent to conduct Network security Drill. The human operator can instruct the Network AI Agent with specific injection policy to work with NDT help construct a dynamic attack-defense verification system in network security drills through NDT and AI reasoning capabilities. The dynamic attack-defense verification system comprise dynamic security attack task agent and dynamic security defense task agent which are responsible security risk attack task and security risk defense task respectively assigned by the network AI agent. The dynamic security attack task agent uses generative AI to automatically generate diversified attack paths, models network topologies with graph neural networks, covers attack stages such as reconnaissance and penetration, and dynamically adjusts strategies via reinforcement learning to simulate the adaptive characteristics of network attacks. The virtual range built based on the NDT can 1:1 map the production environment, supporting simulations of composite scenarios like ransomware chain attacks and supply chain attacks — such as simulating the entire process of Contivirus laterally penetrating to domain controllers through weak passwords. During drills, Human operator can instruct the Network AI Agent to work with the dynamic security defense task agent to automatically deploy virtual environments with vulnerabilities, collect defense response data in real time through NDT, and generate attack path heatmaps and repair suggestions. This capability can further verify emergency response processes, inject real-time threat intelligence to dynamically update drill scenarios, and simulate end-to-end automated deployment, vulnerability injection, and real-time analysis of security drills, enhancing the proactive verification ability of defense systems against real-world threats.

Challenges of Integrating Network Digital Twin and Agentic AI into Network Management In addition to the research challenges in coupling AI and network management specified in , this document also identifies some challenges that need to be considered when integrating agentic AI together with NDT for network operations.

Make IETF developed YANG models AI ready Network management has evolved towards model-driven programmability using YANG (RFC 7950). While YANG provides a rigorous structure for data, it was designed for deterministic software clients rather than the probabilistic reasoning of Large Language Models (LLMs). Consequently, AI agents often struggle to navigate complex YANG hierarchies, leading to "hallucinations" or incorrect configuration logic. The emerging MCP provides a framework for connecting AI models to external tools and data. However, there is currently no standardized way to map YANG primitives to MCP components (Resources, Tools, and Prompts). This gap results in: o Semantic Ambiguity: LLMs lack the contextual hints needed to understand the operational impact of specific YANG leaves. o Interoperability Barriers: AI agents require custom "glue code" for every vendor's unique interpretation of how a YANG model should be exposed to an AI. o Scaling Issues: Massive YANG schemas exceed LLM context windows, requiring standardized methods for sub-schema discovery and pruning.

Trust and Security Multi-Agent Collaborations and interactions can be break down into 4 typical scenarios: o Human operator-> AI Agent-> APIs/Tools/APIs/LLMs In a single-agent scenario, Human operators access services through the network management AI agent. The network management AI agent has multiple functions (fault and optimization), and authentication is required to prevent users from Performing unauthorized actions by exploiting privilege vulnerabilities,e.g., accessing the optimization function interface when they only have fault agent permissions. o AI Agent -> API Services In a single-agent scenario, the network management AI agent triggers tasks automatically based on trace and log information. In some cases, the logging or decision-making process cannot be traced. o Human operator-> AI Agent -> multiple AI Agent Multiple agents may call each other. For example, if a faulty agent A calls an optimization agent B, authentication is required to prevent Manipulating communication channels of agents to influence decision-making processes. o External AI Agent ->AI Agent-> APIs/Tools/APIs/LLMs External AI agents can directly access the network management AI Agent by simulating human operation through interface protocols. For example, a customer AI agent can access the Network AI Agent through a northbound interface protocol such as A2A, MCP. In some case, there might have target flaws in protocols like MCP or A2A; e.g.,consent bypass, context hijacking, etc. Ensuring robust security throughout the entire AI-based network management architecture is essential to prevent unauthorized access and maintain the security of the network infrastructure. The security risk can be break down into the following cases: o External system interacts with AI agent Human Operators or external systems bypass their privileges to operate the Network management AI agent

• Human operators operate through a network management AI agent, but the service scope that the AI agent can handle may exceed the user's authorized privileges, leading to unauthorized access.
• Human operators manipulate business operations using agent delegation and authorization.
• Human operators Coerce intelligent agents to manipulate users into performing covert operations.

o AI Agent Interact with Tools/APIs/LLMs

• Interact with Tools/APIs/LLMs with Privilege Escalation Since network management AI agents rely on LLM for inference when accessing APIs and tools, there is a possibility of malicious injection scenarios where the APIs accessed by the agent exceed the expected scope,e.g.,Using AI to generate execution environments and inject malicious code.
• Interact with Tools/APIs/LLMs without audit When an agent interacts with an API or tool, the logs are currently recorded as system logs, which cannot distinguish between different agents, the logging or decision-making process for specific agent cannot be traced and pose a risk of repudiation.

o Multiple Agent Collaboration and Communication

• Multi-Agent Communication with Privilege Escalation When agents communicate with each other via intent communication and understanding, there is a cascading permission amplification problem,e.g.,AI-generated false information disrupts the reasoning process , leading to Privilege Escalation.
• Interaction with internal AI Agent As a microservice, a network management AI agent can be accessed by other services or AI agents,e.g., Performing unauthorized operations by exploiting authentication vulnerabilities,which poses a risk of Privilege Escalation.

Protocols between Agent and Agent/Human operator/Tools

High Risk Operations "network change" in the network management field refers to the modification, adjustment, or configuration of physical or logical resources in the existing network. Because these operations directly affect the continuity and stability of existing network services, even minor unauthorized access or errors can lead to large-scale network outages. In case of high risk operation, the following measures should be taken into account. o Regardless of whether the initiator of the operation is a human or an agent, each network change request must be re-authenticated. o For operations involving "bulk deletion" or "core route changes", the system must require approval from a second high-privilege account before issuing the command or enforcing the policy. o High-risk configurations should be tested on a very small scale first. The system should automatically monitor the indicators, and if any abnormalities are detected, the configuration should be automatically rolled back within milliseconds.

The Timeliness Requirements of Collaboration For real-time network operation and maintenance scenarios with high real-time requirements, such as scheduling strategy optimization and critical fault repair, the rapid generation of network optimization decisions is crucial. However AI Agents based on large models adopt a "Token-based" generation and reasoning approach, which is limited by computing power and algorithms, resulting in generally slow reasoning speeds. In addition, the simulation and verification process of NDT further increases decision latency, which leads to long end-to-end decision-making time in complex scenarios and is difficult to meet the real-time requirements of services. Also tasks such as fault diagnosis, complaint handling, and user experience improvement often have strict time constraints. For example, if a fault is not resolved within a set time, it can trigger an escalation of the complaint, requiring efficient collaboration among multiple agents. In addition, In a network management environment, you might need agents to subscribe to real-time network alarms or telemetry events. However Google-initiated A2A protocol primarily follows a task-oriented "request-response" model and doesn't support a native pub/sub or event-driven architecture. To improve decision efficiency, continuous efforts are needed in lightweight NDT modeling algorithms, optimizing large model reasoning frameworks (such as quantization technology and parallel computing), and deploying high-performance AI acceleration hardware.

Collaboration Reliability Fault diagnosis and complaint handling in the network management field are complex tasks, typically involving 10 to 20+ fields for one single message exchange between two AI Agents and requiring a high level of expertise. In addition, Reliable task collaboration is incomplete and not sufficient for network management field, e.g., Google-initiated A2A protocol doesn't defined handling strategies for task rejection, missing information during task collaboration, and failure to achieve task objectives. also In network management area, data sources can be diverse and heterogeneous, leading to potential issues such as data inconsistencies, missing, or outdated data. Poor-quality data may result in inaccurate AI predictions and decisions. For example, if incorrect or outdated network configuration data is provided, the model may provide incorrect repair advice when diagnosing network incidents or faults, it may suggest checking an non-existing interface. Ensuring that data is properly cleaned, validated, and maintained is a significant challenge in providing reliable inputs for AI-driven network management.

Agent Observability,Intervention and Control Network operations are increasingly autonomous with the growth of network management Agent applications at the network level and service level. Since AI native operations may be non-deterministic, when network management agents misbehave or deviate from what Agents are expected to do, Current AI safety technologies, often referred to as "AI guardrails" are introduced to constrain the behavior of AI agents within operational and compliance boundaries, prevent AI from producing harmful results or taking wrong actions, e.g.,escalate a decision to a human for a high-risk network operation, defend against malicious attacks,e.g., prompt injection. These guardrails typically operate at the input/output/pre-action filter level or through static boundary alignment. However as AI systems are increasingly integrated into autonomous workflows and critical infrastructure, these static measures are proving insufficient for the full operational lifecycle, they often cannot detect, interrupt, and rollover from unanticipated behaviors. Network operators usually lack an equivalent infrastructure for human oversight or to provide continuous, monitoring of an AI system’s internal logic or its long-running execution paths that match the speed and scale of the network management Agent applications, e.g., network failure or security risk is hard to detect and control, occurring at machine speed. When a violation is suspected, there are currently no standardized protocols for intervention (e.g., immediate task suspension) and recovery (e.g., reverting to a last known safe state or undoing a series of autonomous actions that introduce substantial operational risk) mechanisms. In non-deterministic environments, the lack of human oversight and human-AI semantic intent exchange hinder timely risk mitigation and state recovery during boundary violations by agents.

Agent Benchmarking vs Model Benchmarking Current AI evaluation frameworks are primarily designed for "one-shot" Large Language Models (LLMs) that provide static responses to isolated prompts. However, the industry is shifting toward AI Agents: systems capable of reasoning, planning, and using tools to execute complex, multi-step workflows. Traditional benchmarks (e.g., HumanEval) focus on outcome-only scoring and static knowledge. These fail to measure the core competencies of agents, such as long-horizon planning, tool-use proficiency, and the ability to adapt to environmental feedback. As agents increasingly manage production-level tasks—often involving millions of tokens and hundreds of tool calls—there is a critical need for standardized metrics that evaluate the process of task completion, not just the final result. o Sequential Complexity: Unlike single-turn benchmarks that check one answer, agents require multi-step task completion where each step relies on the success of the previous one. This involves massive workflows (e.g., ~90 tool calls) that traditional benchmarks don't measure. o Environmental Interaction: Agents must interact with live environments. Evaluation shifts from checking if code passes a test to whether a system can select, sequence, and execute tools in realistic, dynamic settings. o Adaptive Planning: Benchmarking must move beyond static knowledge retrieval to evaluate how systems build and adapt plans over long horizons, especially when faced with resource constraints or unexpected disruptions. o Process vs. Outcome: Traditional "right or wrong" scoring is insufficient. Agentic evaluation requires process metrics such as tool success rates, context retention, and multi-turn coherence to determine how an agent reached its conclusion.

Security Considerations The security consideration from apply here. In addition, the following architectural risks need to be considered: o Memory Poisoning：If the AI/ML models used by the network AI Agent or NDT are compromised or poisoned with malicious/fake data,they could begin making incorrect or malicious decisions. Robust checks and validation are necessary to ensure the integrity of these models. Session isolation or memory access authentication is also required to mitigate such risk. o Misuse of Tools： When network AI Agent interacts with tools, Deceptive prompts or commands can be introduced. Tool access verification, tool monitoring, log tracking of AI tool usage are required to mitigate such risk. o Privilege Compromise：When human operators interact with Network AI Agent or Network AI Agent interact with tools/APIs/LLM, unauthorized actions might be performed by exploiting privilege vulnerabilities. Fine-grained permission control, dynamic access verification, and role change monitoring are required to mitigate such risk. o Resource Overload：When Network AI Agent interact with tools/APIs/LLMs, system failures might be caused by exploiting resource-intensive features. Deployment of resource management controls to limit high-frequency task requests from agents is required to mitigate such risk. o Cascading Hallucinations：In case of multi-agent collaboration or communication, AI-generated false information might disrupt the reasoning process.Output verification, secondary verification of AI-generated knowledge are required to mitigate such risk. o Intent Breaking &Goal Manipulation：When external AI Agent or human operators interact with the network AI Agent, reasoning through agent planning capabilities might be manipulated. Planning verification, managing reflection processes, goal consistency protection are required to mitigate such risk. o Misaligned & Deceptive Behaviors：When Network AI Agent interact with tools/APIs/LLMs, harmful operations might be performed by exploiting reasoning vulnerabilities. Manual confirmation of high-risk operations, logging, monitoring, and deception detection are required to mitigate such risk. o Repudiation & Untraceability：In case of multi-agent collaboration or communication, the logging or decision-making process might not be traced. Logging & cryptographic signature, cryptographic verification are required to mitigate such risk. o Identity Spoofing & Impersonation：When Human operators interact with the network AI Agent or in case of multi-agent collaboration or communication,unauthorized operations might be performed by exploiting authentication vulnerabilities. Comprehensive identity verification, trust boundary control, and continuous monitoring are required to mitigate such risk. o Overwhelming HITL(Human In The Loop)：In case of multi-agent collaboration or communication,fatigue auditors attack might take place. Developing advanced human-machine interaction frameworks and adaptive trust mechanisms are required to mitigate such risk. o Unexpected RCE & Code Attacks：When the network AI Agent interacts with tools/APIs/LLMs, Using AI to generate execution environments and inject malicious code might take place. Restrict AI code generation permissions, sandbox isolation, and manual review of generated code are required to mitigate such risk. o Agent Communication Poisoning：In case of multi-agent collaboration or communication, communication channels of agents might be manipulated to influence decision-making processes. Message authentication, communication verification, implementing multi-agent authentication mechanisms. o Rogue Agents in MAS：In case of multi-agent collaboration, there might be malicious or compromised agents. Restricting the autonomy of agents and conducting regular AI testing are required to mitigate such risk. o Humans Attacks on MAS：In case of multi-agent collaboration, business operations might be manipulated using agent delegation and authorization. Restricting the delegation mechanism, implementing AI agent identity authentication, and isolate tasks in segments are required to mitigate such risk. o Human Manipulation：When Human operators interact with the network AI Agent, AI agents might be coerced to manipulate users into performing covert operations. Safety guardrails, content moderation, output content detection are required to mitigate such risk. o Insecure Inter-Agent Protocol Abuse：In case of multi-agent collaboration, there might be target flaws in protocols like MCP or A2A; e.g.,consent bypass, context hijacking, etc. Strong authentication,data validation, restricting delegation to scoped function,logging agent and tool invocations and encrypting communications are required to mitigate such risk. o Supply Chain Compromise：In case of multi-agent collaboration, Vulnerable, malicious, outdated, harmful components might be included into the agent. Digital signatures of SBOMs (AI, Agent),applying version control, chaining authentication, environment isolation are required to mitigate such risk. o Lifecycle security: The entire management lifecycle of the network AI agents and the NDT—from initial deployment and configuration to updates and decommissioning—must be secured against unauthorized access and manipulation.

IANA Considerations This document has no requests to IANA.

Potential Items for IETF Standardization The following items were felt to be good starting points for IETF work:

• OpenTelemetry protocol extension Enabling network behavioral assessment through analysis of observed operational network data (logs, metrics, traces, etc.).
• Human and Agent Interaction protocol for human to control agents, handle agent escalation requests and intervene.
• Nature Language Interaction protocol to ensure both The accuracy and efficiency of structured data for deterministic tasks and natural language interactions for understanding and handling uncertain or ambiguous tasks,e.g., using semantic information transfer or agent prompt language template for Agent to Agent Communication.
• Semantic Metadata: Utilizing YANG Model to provide the Metadata related to Context or Constraint information required by LLMs.

Conclusion This document outlines an architectural framework for AI-driven network operations that integrates a Network Digital Twin (NDT) with Agentic AI. By establishing an Autonomous Domain driven by a Hybrid Agent System and supported by an Agent Fabric, this architecture evolves network management from traditional deterministic automation scripts to flexible, goal-driven, and intent-based operations. While the architectural components and multi-agent collaborative workflows detailed in this document offer a concrete blueprint for autonomous networks, several key challenges remain to be addressed before full production-level deployment can be realized. These include rendering IETF-developed YANG models AI-ready via protocols like MCP, establishing robust trust and security frameworks across multi-agent environments, providing agent observability , intervention and control, and designing advanced evaluation benchmarks that assess task-completion processes rather than just static model outcomes.

Acknowledgements This work has benefited from the discussions of NMRG interim meeting on Agentic AI. Thanks Laurent Ciavaglia, Chris Janz, Pedro Garcia Parra, Brad Peters, Kevin McDonnell, Giuseppe Fioccola, Richard Kilmurray, Dmytro Gassanov, and Chong Feng for wonderful comments and discussion on Agentic AI for Network Operation.

Informative References Intent and Intent-Based Networking are taking the industry by storm. At the same time, terms related to Intent-Based Networking are often used loosely and inconsistently, in many cases overlapping and confused with other concepts such as "policy." This document clarifies the concept of "intent" and provides an overview of the functionality that is associated with it. The goal is to contribute towards a common and shared understanding of terms, concepts, and functionality that can be used as the foundation to guide further definition of associated research and engineering problems and their solutions. This document is a product of the IRTF Network Management Research Group (NMRG). It reflects the consensus of the research group, having received many detailed and positive reviews by research group participants. It is published for informational purposes. China Mobile China Mobile China Mobile Huawei Orange Orange The application of Digital Twin technology in the networking field is meant to develop various rich network applications, realize efficient and cost-effective data-driven network management, and accelerate network innovation. This document presents an overview of the concept of Network Digital Twin (NDT), provides the basic definitions and a reference architecture, lists a set of application scenarios, and discusses such technology's benefits and key challenges. This document is a product of the Network Management Research Group (NMRG) of the Internet Research Task Force (IRTF). This document reflects the consensus of the research group. It is not a candidate for any level of Internet Standard and is published for informational purposes. Daejeon University DONG-EUI University Huawei Everything OPS This document outlines the key objectives of introducing Agentic AI to the field of network management and highlights the fundamental issues with existing technologies that must be addressed to achieve these goals. It emphasizes the necessity for relevant groups within the IETF/IRTF and presents the core technological areas requiring standardization. The aim of Agentic AI is to facilitate a paradigm shift in which multiple autonomous AI agents collaborate to fully automate network operation, management and security. CMCC Telefonica Huawei Ciena This document defines a YANG Module for the network incident lifecycle management. This YANG module is meant to provide a standard way to report, diagnose, and help reduce troubleshooting tickets and resolve network incidents for the sake of network service health and probable root cause analysis. Network telemetry is a technology for gaining network insight and facilitating efficient and automated network management. It encompasses various techniques for remote data generation, collection, correlation, and consumption. This document describes an architectural framework for network telemetry, motivated by challenges that are encountered as part of the operation of networks and by the requirements that ensue. This document clarifies the terminology and classifies the modules and components of a network telemetry system from different perspectives. The framework and taxonomy help to set a common ground for the collection of related work and provide guidance for related technique and standard developments. Ciena Old Dog Consulting Swisscom Huawei Huawei Technologies This document sets out some terms that are fundamental to a common understanding of network fault and problem management within the IETF. The purpose of this document is to bring clarity to discussions and other work related to network fault and problem management, in particular to YANG data models and management protocols that report, make visible, or manage network faults and problems. Huawei Cisco CMCC Telefonica Huawei This document provides an overview of the issues associated with the deployment of the observability, intervention, and control of autonomous agent pipelines in large-scale heterogeneous network environments. The term "Intervention and Control" is used to describe a set of automated and human-initiated mechanisms that guarantee the capability to observe, constrain, correct, and terminate Autonomous agents at any point, for any reason, irrespective of their level of autonomy under which it operates, to ensure resilience, recovery, and operational continuity. The set of enabled observability, intervention and control reflects operator service offerings to ensure that autonomous operations can be stopped, or safely redirected when required and is designed in conjunction with agent to agent, agent to tools, agent to human interaction and service and network policy. This document also identifies several key areas that the Agent Observability, Intervention and Control group will investigate to guide its architectural and protocol work and associated documents. University of Luxembourg and Inria Independent 3NLab Belgium Research Center Canada Post Digital Railway (DSD) at Deutsche Bahn This document is intended to introduce the challenges to overcome when Network Management (NM) problems may require coupling with Artificial Intelligence (AI) solutions. On the one hand, many difficult NM problems still lack good solutions, or existing approaches come with significant limitations. Artificial Intelligence may help produce novel solutions to those problems. On the other hand, due to the high computational costs of AI solutions and stringent data privacy constraints, the distributed execution of AI workloads has become paramount. Consequently, networks must be operated efficiently to sustain these distributed processing requirements. To identify the right set of challenges, the document defines a method based on the evolution and nature of NM problems. This will be done in parallel with advances and the nature of existing solutions in AI in order to highlight where AI and NM have already been coupled together or could benefit from a closer integration. So, the method aims at evaluating the gap between NM problems and AI solutions. Challenges are derived accordingly, assuming that solving these challenges will help to reduce the gap between NM and AI. This document is a product of the Network Management Research Group (NMRG) of the Internet Research Task Force (IRTF). This document reflects the consensus of the research group. It is not a candidate for any level of Internet Standard and is published for informational purposes. This document specifies the IP Flow Information Export (IPFIX) protocol, which serves as a means for transmitting Traffic Flow information over the network. In order to transmit Traffic Flow information from an Exporting Process to a Collecting Process, a common representation of flow data and a standard means of communicating them are required. This document describes how the IPFIX Data and Template Records are carried over a number of transport protocols from an IPFIX Exporting Process to an IPFIX Collecting Process. This document obsoletes RFC 5101. This document defines a YANG data model and associated mechanisms enabling subscriber-specific subscriptions to a publisher's event streams. Applying these elements allows a subscriber to request and receive a continuous, customized feed of publisher-generated information. This document describes a mechanism that allows subscriber applications to request a continuous and customized stream of updates from a YANG datastore. Providing such visibility into updates enables new capabilities based on the remote mirroring and monitoring of configuration and operational state. This document defines the BGP Monitoring Protocol (BMP), which can be used to monitor BGP sessions. BMP is intended to provide a convenient interface for obtaining route views. Prior to the introduction of BMP, screen scraping was the most commonly used approach to obtaining such views. The design goals are to keep BMP simple, useful, easily implemented, and minimally service affecting. BMP is not suitable for use as a routing protocol. The Network Configuration Protocol (NETCONF) defined in this document provides mechanisms to install, manipulate, and delete the configuration of network devices. It uses an Extensible Markup Language (XML)-based data encoding for the configuration data as well as the protocol messages. The NETCONF protocol operations are realized as remote procedure calls (RPCs). This document obsoletes RFC 4741. [STANDARDS-TRACK] This document describes an HTTP-based protocol that provides a programmatic interface for accessing data defined in YANG, using the datastore concepts defined in the Network Configuration Protocol (NETCONF). YANG is a data modeling language used to model configuration data, state data, Remote Procedure Calls, and notifications for network management protocols. This document describes the syntax and semantics of version 1.1 of the YANG language. YANG version 1.1 is a maintenance release of the YANG language, addressing ambiguities and defects in the original specification. There are a small number of backward incompatibilities from YANG version 1. This document also specifies the YANG mappings to the Network Configuration Protocol (NETCONF).

Network Element AI Agent and Network AI Agent Collaboration As a supplementary deployment case, this section explores the collaboration mechanism between network element AI Agent and Network AI Agent. Network devices collect information from multiple dimensions, including flow information, configuration, events, alarms, logs, dynamic topology and routes, and device status (including CPU, memory, and hardware health). With large amount of data collected to the domain controller for analysis and processing, the data accuracy is very limited and therefore it is hard to determine the service impact within 1 minute. In addition, it usually require multiple step interaction, complex task management with various different data types or data sources. To address those challenges, the network AI Agent can delegate massive data analysis and processing to distributed AI Agent in each network element, e.g., a) only allow distributed AI Agent export processed analytic data to help establish global view of network observability. b) or export key network fault information for Network AI Agent for further investigation the root cause of the problem. For the former case, routing protocol specific fault data such as BGP Status Changed, OSPF Neighbor state changes, IS-IS Adjacency Changed data or hardware related fault data such as Optical fail, Physical Port down can be collected and using pre-trained LLM model with expert experience to match fault pattern and invoke corresponding routing protocol troubleshooting MCP tools and finally root cause. In addition, it allows network maintenance engineer using nature language interface to look up troubleshooting information or it allows Network AI Agent or Task Agent at the network element using MCP interface to invoke tools from MCP server within the network element.

Network Element AI Agent and Network AI Agent Collaboration Usage Example MCP | || | | | AI Model| |Client| | | | Server | || | | +---------+ +------+ | | +------------+ || | +-----------------------+ +--------------------+| | | +----------------------------------------------------+ Smart Network Element ]]>

AI Agent Registration and Team formation this section provides an illustrative realization example demonstrating how agent registration could help with the dynamic team formation when a specified network intent is requested to be handled.

AI Agent Registration and Team formation Usage Example | | +-----------+ | |Registration| | |---------------+----------------+---------+-> | | Register Register Register | | Center | | | | | | +------------+ | +------+----+ +------+-----+ +-----+----+ | | | Fault | |Optimization| | Change | | | | Task Agent| |Task Agent | |Task Agent| ...| | +-----------+ +------------+ +----------+ +----------------+ +-----------+ | Incident | | Center | +----------------+ +-----+-----+ |Agent Fabric | Incident Notif | +------------+ | | Registry based | | Agent | | Intent +-----V-----+ Discovery | | | | Objective | Network +--------------------+->Registration| | Human ---------> AI Agent | | | | | Operator | | | | Center | | +-----+-----+ | +------------+ | Policy Definition(Obj+Incident+Action) | | Task Planning | | Agent Team Formation | | +---------------+----------------+ +----------------+ | | | +-----+-----+ +----+------+ +----+------+ | | | | | | |Task Agent1| |Task Agent1| |Task Agent1| +-----------+ +-----------+ +-----------+ ]]>

The Agent Fabric at the network level provides agent registration for both network AI agent and associated task agents. The following steps are performed to provide event-driven AI Agent Team formation within the Agentic AI network management architecture:

Step 1:

Human Operator pre-provision user intent which comprises objective, incident list and corresponding action list.

Step 2:

Network AI Agent receieves user intent and generate corresponding policies which comprise objective, incident list and corresponding action list. In addition, Network AI Agent subscribe corresponding incidents from incident center.

Step3:

Network AI agent generate task planning based on objective and then discover matched task agent lists based on planned task from registration center within the agent fabric.

Step4:

Upon receiving incident from the incident center, network AI agent assign the tasks to task agents corresponding to specific incident received.

Step 5:

Network AI Agent distribute task to corresponding task agents and complete task agent team formation.

Agent to Agent Communication Security This section provides an informative usage example to illustrate how to secure an Agent to Agent communication via protocols such as A2A.

Agent to Agent Communication Security Usage Example Agent | | | | AI Agent <---------------------+ | | | | | +-^------^--+ Authentication | +--+->Registration| | | | | Handshake | | | | | | | Secure | Stateful | | | | Center | | | Task | Monitoring | | | +------------+ | Negotiation| | | | | | | | | Agent Identity | | | +-------------+| | +-V------V--+ Discovery +--+--+-> Agent || | | Task +------------------------+ | |Authorization|| +--------> AI Agent <---------------------------+-> Server || +-----------+ Authentication | | || Handshake | +-------------+| +----------------+ ]]>

The following steps are performed to provide Agent to Agent Communication Security within the Agentic AI network management architecture:

Step 1. Discovery via Agent Card:

The workflow begins when a client agent requests the Agent Card from the remote agent's /.well-known/agent.json endpoint. This JSON file acts as a secure manifest, declaring the agent’s identity, capabilities, and required security schemes.

Step 2. Authentication Handshake:

Before any task is sent, the client must fulfill the authentication requirements listed in the Agent Card. This typically involves an OAuth 2.0 flow where the client obtains a JSON Web Token (JWT) to prove its identity and permissions.

Step 3. Secure Task Initiation:

Communication is established over HTTPS/TLS. The client sends a tasks/send request using JSON-RPC 2.0. The server validates the token and authorizes the specific task based on the client's role.

Step 4. Stateful Monitoring & Feedback:

The task moves through a strictly defined lifecycle (submitted → working → completed). Security is maintained throughout as updates are streamed via Server-Sent Events (SSE) or webhooks, each tied to the unique, authorized Task ID.

Step 5. Artifact Delivery:

Final results (Artifacts) are delivered only after the task reaches a completed state. These are structured objects (text, files, or data) returned to the verified requester, ensuring data integrity and preventing unauthorized access to output.

Changes between Revisions v05 - v06 v03 - v04 v02 - v03 v00 - v01

Contributors Huawei

maqiufang1@huawei.com

CMCC

lizhenqiang@chinamobile.com

Orange Research

lionel.tailhardat@orange.com